Adversarial Learning for Incentive Optimization in Mobile Payment Marketing

Abstract

Many payment platforms hold large-scale marketing campaigns, which allocate incentives to encourage users to pay through their applications. To maximize the return on investment, incentive allocations are commonly solved in a two-stage procedure. After training a response estimation model to estimate the users' mobile payment probabilities (MPP), a linear programming process is applied to obtain the optimal incentive allocation. However, the large amount of biased data in the training set, generated by the previous biased allocation policy, causes a biased estimation. This bias deteriorates the performance of the response model and misleads the linear programming process, dramatically degrading the performance of the resulting allocation policy. To overcome this obstacle, we propose a bias correction adversarial network. Our method leverages the small set of unbiased data obtained under a full-randomized allocation policy to train an unbiased model and then uses it to reduce the bias with adversarial learning. Offline and online experimental results demonstrate that our method outperforms state-of-the-art approaches and significantly improves the performance of the resulting allocation policy in a real-world marketing campaign. CCS CONCEPTS• Applied computing → Marketing.

Marketing Budget Allocation with Offline Constrained Deep Reinforcement Learning

“…Many classic methods use a two-step framework [5,8,19,23,35,36,39], i.e., a response prediction step and a decision making step. Response prediction models include DNN [8,39] or GNN [23,36], while the decision making step can utilize dual method [40], linear programming [8,36], bisection method [39] or control based method [35]. These methods only optimize the immediate reward and fail to capture the long-term effect.…”

“…As these metrics are hard to directly optimize, conventional methods use immediate user responses, like the coupon redemption rate [36], as surrogates. Typically these methods take a two-step framework [5,8,19,35,39]. They first build a response model, which estimates users' immediate responses to different incentives [8,23], then solve a constrained optimization problem to make the budget allocation [5,40].…”

“…Typically these methods take a two-step framework [5,8,19,35,39]. They first build a response model, which estimates users' immediate responses to different incentives [8,23], then solve a constrained optimization problem to make the budget allocation [5,40]. Although this framework scales well, it fails to take the long-lasting effects of a marketing campaign into consideration.…”

A Survey of Deep Causal Models and Their Industrial Applications

“…Incentive allocation relies on treatment effect estimation models to estimate users' purchase probability with different incentives. PCAN [124] was proposed to learn an an unbiased model by leveraging a small set of unbiased data. Speciőcally, a biased network was built to generate unbiased data representation by controlling the distribution difference to a unbiased network.…”

Imbalance-Aware Uplift Modeling for Observational Data

“…In addition, with the flexibility of the design of neural networks, it is easy to realize deconfounding of the uplift modeling on the non-RCT data. Several deep learning based methods (Johansson, Shalit, and Sontag 2016;Yao et al 2018;Yu et al 2021;Ma, Li, and Cottrell 2020;Li et al 2021;Künzel et al 2018;Yao et al 2019;Chen et al 2021;Yao et al 2021) successfully extend the traditional approach to combine with deep learning and achieve improvements on the uplift modeling.…”